Generate first benchmark agent runs for agentos.org leaderboard

[andyk]

We should populate our public registry w/ training and evaluation Runs of the most popular OSS agent x environment pairs. We'll start with 2 agents and one env., and then add more agents and more envs.

We currently have SB3 PPO playing cartpole.
FINE GRAINED TODOS:

• Factor out CartPoleEnv into a common component whose GitHubRepo points to OpenAI's repo.
• add PAPAG agent playing cartpole env.
• add Ilya PAPAG agent playing gym-pong env.
• add SB3 agent playing gym-pong env.

Below are our current shortlists of other agents and envs to benchmark:

Agents/algos

1. SB3
2. Ilya PAPAG PPO agent
3. RLlib
4. Acme
5. RLkit
6. Open-clip

Environments

1. Atari games (one easy, one of the hardest)
2. PyBullet or Mujoco simulator (humanoid? half cheetah? soccer?)
3. Robotic arm simulator
4. Classic control?
5. DeepmindLab?
6. Minecraft?
7. [STRETCH] Hottest NLP (??) or vision benchmarks (ImageNet)

We should publish our code for these runs as new scripts in documentation/demos/ so that they double as helpful learning resources & examples of how to use the AgentOS & PCS APIs.

Note the registry is currently hosted at https://aos-web.herokuapp.com, after #331 it will be hosted at agentos.org/registry.